Multi-Level Floyd-Steinberg Dithering Boosts Adversarial Robustness in Vision Models
There’s this new study on arXiv (2605.23065) that presents a fresh, lightweight method to defend vision foundation models against adversarial attacks using a technique called multi-level Floyd-Steinberg error-diffusion dithering. This method successfully reduces adversarial noise while keeping the semantic accuracy intact. Unlike previous studies that only looked at binary dithering for a single small model on grayscale CIFAR-10, this research evaluates six different tasks, including classification and segmentation, across two model families, DINOv2 and PaliGemma. It also tests against three increasing levels of attacks and an adaptive attacker. The results show that using Floyd-Steinberg dithering with intermediate quantization levels, especially with some post-processing blur, performs as well as or better than all baseline methods.
Key facts
- arXiv preprint 2605.23065 proposes multi-level Floyd-Steinberg dithering for adversarial defense.
- Method is lightweight and model-agnostic.
- Evaluated on six tasks: classification, segmentation, depth estimation, retrieval, captioning, visual question answering.
- Two model families tested: DINOv2 and PaliGemma.
- Three attacks used: PGD, MI-FGSM, SIA.
- Adaptive attacker with straight-through estimator included.
- Intermediate quantization levels with post-processing blur yield best results.
- Prior work limited to binary dithering, grayscale CIFAR-10, and single small model.
Entities
Institutions
- arXiv